Trufo at ICIP 2026: Watermarking and JPEG Trust

In Tampere, Bill (Trufo’s watermark research lead) presented at the 2026 IEEE International Conference on Image Processing (ICIP), on their watermark submission to the JPEG Trust Watermarking Grand Challenge that Trufo’s PawPrint v2024b won. The first half of the presentation focused on the actual PawPrint v2024b implementation, while the second half of the presentation focused on the state of watermarking in the emerging content provenance industry.

The competition, hosted by the JPEG Trust team (ISO/IEC 21617), featured participants from 12 countries, including 8 industry entrants. At ICIP, Trufo was joined by the other winner of the competition, a research team from KETI, a research lab in South Korea, who was able to substantially improve the performance of PixelSeal using semantic embedding teaching.

Trufo’s PawPrint v2024b submission to the competition was a traditional watermark, not the current PawPrint that runs in production today. The v2024b watermark was initially developed in 2024 and modified to match the 100-bit payload that the watermark challenge expects. What is impressive about the v2024b watermark is that not only was it the strongest watermark in academia and industry until the release of PixelSeal by Meta, but it held this position at a time when the SotA had largely transitioned from traditional methods to deep learning methods. Now, in 2026, PawPrint v2024b remains stronger than industry competitors (unless you count fine-tuning PixelSeal), as evidenced by this competition. The full paper is below:

The benchmark used in the JPEG Trust Watermarking Grand Challenge is the first major watermark benchmark balanced around the needs of content provenance, which means a focus on both imperceptibility metrics and on benign durability metrics (e.g. realistic compression and other common edits). The benchmark may not be perfect (and we calibrate our watermarks using a different set of benchmarks and objectives altogether), but it is a good impartial indicator of watermark performance.

Note that the watermarks used in production at Trufo are not PawPrint v2024b, but rather a neural watermark, using a substantially distinct architecture from TrustMark or PixelSeal or SynthID, trained from the ground up.